Domain depth, sharpened by AI.

Quinn is a chat assistant for semiconductor manufacturing data. Built at HP with the product and process engineers who use it. Pilot since Jun'26.

An engineer asks in plain words. Quinn finds the lots, pulls lot history, electrical test, metrology and inspection data, maps pedigree across wafer, module and assembly segments, runs the analysis and renders the maps. Follow-ups stay in the same chat. Any answer can be kept in a shareable report.

The model picks the analysis. Tested code computes every number.

• The model reads a catalog of 55 tools and chooses the tool and its arguments. It sees the prompt, its own calls and a capped summary of each result. Never the full data.
• Every call is checked against the tool's schema before it runs. SQL, statistics and maps are first-party code.
• Model-written code stays fenced. Pandas runs in a sandboxed child process. SQL runs as one SELECT on the read-only local mirror, never on the warehouse.
• The model runs on the company's internal LLM gateway. Data, maps and full results stay on the Quinn host.
• A turn that fails after tools have run returns a partial envelope. --> Completed analysis is never discarded.

Agents

17 agents, 55 tools in the model's catalog. Extraction, analysis, visualisation, tool commonality, correlation, split analysis, lot and spatial localisation, cross-segment pedigree, yield reports, files dropped into the chat.

Statistics

Fisher's exact test with Benjamini-Hochberg FDR for tool commonality. Welch's t and random-effects log-odds for process splits. Lot-clustered Pearson and Kruskal-Wallis for cross-segment correlation. Excess-mass origin verdicts for localisation.

Stack

Python, Streamlit, a DuckDB mirror of the warehouse refreshed nightly, a FastAPI engine sidecar that runs the tool loop, JSON-schema tool contracts. Open-weight Qwen models on Ollama as the local option.

Testing

Routing is graded by replaying a 200-turn prompt catalog through the model. Offline contract tests pin every tool's schema.

Output

Tables, wafer, panel and pick maps, Pareto and fail-rate charts, HTML reports with the CSV behind every table. Daily, weekly and monthly yield monitoring that runs localisation and tool commonality on each signal.

Draw a box on the layout and pick the process script. Every deposit, etch and implant runs over the whole region, and the result opens in KLayout's 2.5D view. The 4 side faces are the 2D cross sections along the edges of the box.

• Same process scripts and layer files as KLayout's 2D XSection package.
• Cut into the model from any face. Make an oxide see-through. Stretch the vertical axis for thin films.
• Trace a net in the layout and it lights up in 3D: metal, plugs and diffusions.
• Process steps run on all cores. Every face is checked against the 2D cross section.

Built

At HP, Oct'26. About 8,000 lines of Ruby on KLayout.

Pictures

A generic SOI test chip drawn for this page: 220nm silicon on 2um oxide, strip and rib waveguides, n and p implants, tungsten contacts, a TiN heater, aluminium. No foundry's process or design.

Can a graph over Singapore's HDB blocks price a block that has never been resold, and which graph works best? Each of the 10.8k blocks is a node, linked to its nearest neighbours and described by 13 layers of nearby amenities: MRT, schools, hawker centres, parks, malls and more.

• GCN, GraphSAGE, GAT, a learned-edge convolution and a Graph Transformer, written from scratch in PyTorch. Against an MLP, LightGBM, kNN and linear baselines.
• 280 runs: 18 models, a temporal and a spatial split, 4 graph densities, 5 seeds, an amenity ablation. 1h50min on a DGX Spark.
• LightGBM with coordinates wins everywhere: 0.052 RMSE on log price per m², about 4% error. No neural model beats it on any 5-seed mean.
• Amenities and message passing are substitutes. With no amenity features the GCN beats the same network without a graph on held-out blocks in all 5 seeds. With 13 layers the gain is gone.
• New estates are 8yrs old against 34 for the blocks the models trained on, and sell about 33% higher per m². Every model under-prices them, the more so the more it leans on its neighbours.

Course

NUS CS5284 Graph Machine Learning. A group project, done solo, Oct'26.

Data

241k HDB resales, Jan'17 to Sep'26, from data.gov.sg. Amenities from data.gov.sg (8 layers) and OpenStreetMap (5).

Map

Faint to strong blue: S$2.4k to S$5.6k per m², in 2009 prices. Amber: the 138 blocks with no resale before 2025.

Edges

The learned-edge convolution found estate boundaries by itself. At the same distance, a neighbour on the same street weighs 3.3 times one on another street.

Finding

The new estates a graph should help most are the least like their neighbours.

Grasp planners score a grasp through the friction at the fingertips, yet almost always assume one friction value for every object. This project asks whether estimating it from a single photo, before contact, picks better grasps.

• Photo, then a material classifier, then a spread of friction values. Plan at the pessimistic end of that spread, rank the grasps and set the grip force.
• Tested in PyBullet on 10 YCB objects, paired: one constant friction, an oracle that knows the true value, and a 2×2 that separates which grasp is chosen from how hard it is squeezed.
• Knowing the friction raises lift success by 5 points, 9 on slippery objects and 26 on metal. All of the gain comes from grip force. Choosing a different grasp with the true friction does not help.